FICE: Fovea-Inspired Closed-Loop Edge-AI System for Enhanced Capsule Endoscopy Imaging

The wireless feature of a capsule endoscope enables non-invasive gut imaging, but it results in limited power supply, poor imaging quality, and low frame rate. This paper proposes a Fovea-Inspired Closed-Loop Edge-AI (FICE) system to address these constraints without compromising the diagnostic quality. Specifically, we integrate a lightweight transformer-based model for onboard lesion detection and a Gated Corner Proposal Network for precise region-of-interest (ROI) localization. Once lesions are flagged, the system captures and transmits high-resolution ROI images while maintaining low-resolution frames elsewhere, effectively conserving energy and bandwidth. Key findings using the HyperKvasir and Kvasir-SEG datasets demonstrate a lesion detection recall and precision of 99.88% and 99.40%, respectively, along with a mean Intersection over Union (mIOU) of 0.804 for bounding box localization, while maintaining a processing speed of 9.58 fps at 320-pixel density under a maximum data transmission rate of 1248 KBps. The frame rate and pixel density are improved by five and three folds, respectively, compared to the state-of-the art capsule endoscope. This closed-loop approach mimics the human eye's fovea and saccadic movements, balancing the need for detailed imaging of critical areas with overall resource constraints.